taekbae/hanabi-zsc-llm-game-hidden

Hanabi ZSC — hidden states from LLM-vs-LLM games (Llama-3.1-8B family)

Per-turn residual-stream activations recorded while three Llama-3.1-8B-based agents played 240 self-play Hanabi games each, in a zero-shot-coordination (ZSC) setting where the partner's hint convention is latent and must be inferred from the log. Released so that probes and read-out analyses can be redone without any GPU forward pass.

folder	model	games	files	size
llama31/	NousResearch/Meta-Llama-3.1-8B-Instruct	240	480	~70 GB
tulu3/	allenai/Llama-3.1-Tulu-3-8B	240	480	~70 GB
hermes3/	NousResearch/Hermes-3-Llama-3.1-8B	240	480	~70 GB

One file per (game, seat): A_c{config}_e{episode}_s{seat}.npz. Two seats per game, so 480 files per model. A marks the condition: the receiving LLM gets no injected sentence and no probe value (the "raw" cell); hints are emitted by a scripted policy that follows the hinter's own convention.

The task

Two players, mirror-world Hanabi. Each holds a private convention with two axes: focus — when a hint marks two or more cards, does it mean the leftmost or the rightmost marked card; kind — does a colour hint mean "play" (and rank "discard") or the reverse. A convention is drawn per game from the seed, kept fixed for the whole game, and never stated in the prompt; the prompt says only that the partner follows a consistent unstated rule. An agent can infer the partner's convention only from the hints and reactions that accumulate in the public log.

Seeds: seed = 2,300,000 + config × 100,000 + episode, config 0–3 = (LEFT,COLOR)·(LEFT,RANK)·(RIGHT,COLOR)·(RIGHT,RANK), episode 3000–3059 (60 boards × 4 configs = 240 games). Each seat sees the same game from its own side: actor labels, the hand it can see, and the "your rules" header all swap.

What is in each file

Arrays (fp16 for activations, int32/int64 for indices):

array	shape	what
turn_hidden	(turns, layers, d)	last token of the decision prompt, no read-out question
turn_q_hidden	(turns, layers, d)	last token of decision prompt + read-out question, run as a separate forward pass
hist_hidden	(Σ history lines, layers, d)	end-of-line token of every history line, recorded per turn (not deduplicated)
hist_turn, hist_event_idx	(Σ history lines,)	which turn and which event index each history-line row belongs to
turn_no, turn_hist_len, turn_n_token	(turns,)	turn number, number of events visible, prompt length
token_ids, token_offset	flat	the exact token sequence of each turn's prompt
meta	JSON string	seed, config, episode, seat, both seats' assigned conventions, model, tag, question text, argv, git commit, timestamp, host

layers = 33 for the 8B models (embedding output + 32 blocks), d = 4096. Activations are the residual stream (output_hidden_states) at the selected positions.

The read-out question appended for turn_q_hidden is:

Consider P1's hints that marked two or more slots and what was played after them: is P1's focus the LEFTMOST or the RIGHTMOST marked slot?
Does P1 use COLOR or RANK hints to mean play?
ANSWER:

record_*.pkl in each folder is the game record the hidden states were taken from (per-turn state, event list, both conventions, per-decision logs). It is a Python pickle of the producing repository's own structures; the npz files are self-contained for probing, the pickle is there for anyone who wants to recompute which events count as evidence.

Intended use

Train a linear probe on activations to test whether the partner's convention is linearly readable at a given position, then compare positions: the summary position (turn_hidden / turn_q_hidden, one vector per turn) against evidence positions inside the history (hist_hidden). In the source project the summary read gives the model's own aggregate, while reading evidence lines and summing outside the model gives an upper bound on "the pieces are there but the model does not aggregate them".

Labels come from meta: partner_conv is the convention of the seat that hinted to this seat — that is the quantity a probe should predict. own_conv is this seat's own convention (stated in its prompt header, so trivially readable; useful as a control).

A minimal read:

import numpy as np, json
z = np.load("llama31/A_c0_e3000_s0.npz")
meta = json.loads(str(z["meta"]))
X = z["turn_q_hidden"].reshape(len(z["turn_no"]), -1)   # (turns, 33*4096)
y = meta["partner_conv"]          # e.g. ["FOCUS_LEFT", "COLOR_PLAY"]

Caveats

• These are activations of specific released checkpoints of Llama-3.1-8B-Instruct, Tulu-3-8B and Hermes-3-Llama-3.1-8B, in bf16 on NVIDIA GPUs. Numerics differ across hardware and dtype; do not mix with activations collected elsewhere without checking.
• Games are self-play (both seats the same model), so the two seats of one game are not independent samples — group by game when splitting train/test.
• hist_hidden repeats each history line once per turn on purpose (the "line activations do not change across turns" argument is not assumed, so it can be checked).
• meta keeps the producing machine's hostname and the command line, as provenance.

Provenance

Produced by env2_hidden_store.py in the project's zsc_final tree (research code, commit hash in each file's meta.git), 2026-09-21/22, three GPUs, sharded by seed. Companion release with the run logs and probe checkpoints of the same project: taekbae/hanabi-zsc-runv1-llama.

Model licences apply to derived artefacts: Llama 3.1 Community License for Meta-Llama-3.1-8B-Instruct and Hermes-3-Llama-3.1-8B, and Llama 3.1 Community License plus the Ai2 terms for Llama-3.1-Tulu-3-8B. This dataset contains no human-generated content: all game states are synthetic and generated from fixed seeds.